The Value of Explainability in Artificial Intelligence

The Value of Explainability in Artificial Intelligence

AI applications are finding a role in many business processes. Machine learning algorithms add speed, precision and automation to enable companies to drive improved efficiency and performance in across a wide range of discrete tasks. Explainable AI though does all that and more. It adds insights to the equation that enable businesses to move beyond automation to actionable intelligence.

Actionable intelligence creates business gains and prevents losses on the top line, not just efficiency or better precision. It provides decision systems and business leaders with the intelligence to discover, see and act on key insights that are generated every day. Insight driven businesses that master explainable AI will increase speed to insight and strengthen competitive advantage.

Millions of online purchase transactions flow through algorithms predicting, based on past transactions, that a new transaction has a higher confidence of fraud. However, some transactions are just above the normal vs. fraudulent threshold. Before this potentially valid transaction is denied, upsetting the customer, it would be nice to know Why it is being flagged by the algorithm as fraud.

Upon review of an explanation, a business user can determine that this is a new device with a higher than average purchase volume, but other identification factors indicate it’s a legitimate, repeat customer. Without knowing the factors associated with the prediction, the customer would be denied, driving them to go elsewhere. Multiply this across thousands of customers and you have a real problem. An explainable algorithm that can produce the factors associated with each prediction, can prevent this from happening.

Millions of authentication instances occur each day. Over time user ID’s and passwords fragment across devices and applications. A customer tries to log-in but forgets his password and after several attempts, is prevented from accessing his account. Frustrated, the customer gives up on logging in and doesn’t come back to use the application any longer, selecting another alternative.

Device ID data, however, from other interactions with the same customer suggest this is a known user that you can offer a more seamless set of verification steps to, using other known information. Explainability enables an adaptive approach, by exposing all of this data in real-time and at scale to authentication decisioning systems.

Thousands of customers or subscribers are predicted to churn each month for many large B2C enterprises but without knowing why, it is impossible to take action to prevent the customer from leaving with specific offers that address their issue. When hundreds of customers a day are predicted to churn, how do you act on each one effectively without knowing why? Explainability enables an actionable response for each and every type of churn cause.

The value of an explainable machine learning prediction like those reflected above, are sometimes hidden underneath the covers. It’s not hard for people to appreciate the logic of explainability, but it is not always clear how explainability translates to insights and then business gains or loss prevention.

For example, it doesn’t take much these days for a customer to leave you. One bad experience vs. hundreds of good ones is sometimes all that it takes to lose a customer. And what if a customer is a highly valuable one? This is where the cost of not having explainable insights at speed and scale starts to add up.

In their article “AI Is Going to Change the 80/20 Rule,” published in the Harvard Business Review in 2017, authors, Clayton Christiensen, Michael Raynor and Rory McDonald comment on the power of the Pareto Principle, or the 80/20 rule. They comment that high-performance organizations “are still inspired by his (Pareto’s) 80/20 principle, the idea that 80% of the effect (sales, revenue, etc.) come from 20% of causes (products, employees, etc.).”

They go on to point out that “greater volumes and variety of data guarantee that algorithms get the training they need to get smarter. Digital networks consequently become Pareto platforms that transform vital vectors of variables into new value.

Novel workplace analytics, for example, mean more organizations can more readily identify the 20% of employees contributing 80% of value to a product, process, or user experience. Ongoing digitalization of business processes, platforms, and customer experiences similarly invites creative Pareto perspectives: What 20% of the platform upgrade creates 80% of its impact? What 20% of customer experience evokes 80% of delight or distaste? Serious C-suites want those data-driven questions algorithmically addressed.”

How do you optimize your business if you don’t know the cause, or key drivers, behind a prediction? Without knowing “why” an algorithm is making a decision, not only can that severely constrain your business operationally, but it’s a missed opportunity for understanding the 20% factors that cause 80% of the best (or worst) outcomes.

We see this right now across fraud, identity and marketing domains inside major enterprises on a regular basis. Here are 5 ways explainable machine learning algorithms create speed to insight that can have an immediate and outsized impact on your business.

Insight driven businesses can discover information about customers, transactions, products, geographies, or even vast amounts of social media driven perspectives it didn’t previously know.

For example, uncovering the fact that certain customer segments who buy one product, don’t buy another, vs. other segments that do is a pretty significant discovery. What are the causal factors for this? Do certain product features appeal to one group and not another? Is it an economic issue such as price point or income – or the combination of both?

Knowing the cause will enable the business to change its offers to the non-buyer groups or only focus offers to buyers who look like ones who bought the additional product. One approach could increase sales, the other reduce cost. These basic decisions can only be supported if algorithms provide explainability.

In the article mentioned above, another example was provided. “For instance, one multibillion-euro industrial equipment company with over 2,000 SKUs determined that less than 4% of its offers were responsible for one-third of sales and roughly half of profitability. But extending the analysis to include service and maintenance revealed that roughly 100 products were responsible for over two-thirds of profitability. That pushed the firm to fundamentally rethink pricing and bundling strategies.”

Seeing new anomalous instances of behavior or activity is a way to notice an emerging trend. The data supporting a new anomalous event, such as a fraud attempt that has never before been seen, and slipped through, will provide the foundation for spotting an emerging trend.

What characteristics define it? Will another event with similar characteristics occur and if so how often? If it does occur again are their characteristics about this event that make it preventable from a business perspective? Emerging trends can be dangerous or positive – either is a missed opportunity to prevent significant losses with fast action, or realize a new. A large scale hardware manufacturer for example monitors not only when a server is experiencing a likely “failure” event, but is examining why, in order to enable fast, preventative action based on the known factors associated with each failure prediction.

Patterns can be subtle. The cause and effect of weather, political or sporting events, economic changes, the change or action of a competitor, negative publicity, etc. can correlate to a broad range of impacts in ways that are not seen.

Explainable algorithms can see these correlations and reveal the main drivers behind them. If we see a certain event occur, what patterns in behavior do we understand from the past to predict and act on the likely to occur effects of that event happening again.

For example, for a large retailer, during periods of very hot temperatures, sales of pool toys spiked in certain rural regions of the country where there were nearby lakes vs. more urban areas with pools. This pattern allows inventory to be shifted to the rural store locations when temperatures are forecasted to exceed a certain level. Explainable AI can reveal these kinds of patterns.

Knowing the cause or reasons that explain a predicted behavior allows for action. Earlier, we mentioned likelihood to churn as one example. Not knowing why, paralyzes a company from appropriately preventing the churn from happening. The same applies to product recommendations.

Why does a shopper buy certain products on Tuesdays, and others on Thursdays? Knowing why can enable email notifications to customers that are relevant and personalized – such as the new shipment of their favorite brands of bread and dairy products just arrived, which has typically occurred on Tuesdays.

Explainable algorithms can reveal this basic information and causal drivers if fed the right information. In a world where relevancy is critical, explainability makes algorithmic predictions actionable and customer communication relevant.

Innovation can be constantly fed from existing customers and how they use what you provide. If high value customers demonstrate common characteristics of usage or lack thereof, perception, inactivity, neglect, reduction in volume, etc. This can lead directly to innovations that are fueled by why customers behave or think as they do.

This is common in software for example. If someone is accessing their SaaS software less and less, after having been a heavy user, are there characteristics of their use of the product that began to frustrate them? Did a change in software features cause problems for them? Analyzing and clustering usage can reveal opportunities for innovation – creating a new feature that makes something easier for example.

“Super Peretos” is another concept that the vital few (80/20) becomes the vital fewer (95/5). According to the authors referenced previously, “Analytically aggressive firms increasingly see Pareto proportions closer to 10/90, 5/50, 2/30, and 1/25. Depending on how rigorously the data is digitally sliced, diced, and defined, 1/50, 5/75, and, yes, 10/150 Paretos emerge. Pareto’s “vital few” becomes a “vital fewer.”

They go on to explain that; “Extreme distributions transcend and dominate the industry. Fewer than 10% of drinkers, for example, account for over half the hard liquor sold. Even more extreme, less than 0.25% of mobile gamers are responsible for half of all in-game revenue.

Clearly identifying and cosseting the “super-Paretos,” however, doesn’t go analytically far enough; market and market growth demand that those descriptive statistics lead to predictive and prescriptive statistics. In other words, turn those data sets into “training sets” for smart algorithms.”

You can see the impact that speed to insights can have when applied in the context of the Pareto principle, let alone Super Pareto principle. Speed and insight combine together to drive a business to be a true insight driven one that can have explosive growth. Isn’t it worth it, therefore, to invest in Explainable AI algorithms running on massive training data sets looking for these insights all the time? Insight driven businesses think so and do so.